[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-intern-decision-qwen3-5-multimodal-decision-model":3,"topics-all":44,"news-related-48726cf0-6f0c-4b09-841e-06aa8b0ffcfd":63},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":30,"news_slug":37,"published_at":38,"created_at":39,"modified_at":40,"is_published":41,"publish_type":42,"image_url":14,"view_count":43},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","上海 AI Lab 发布 Intern-Decision 多模态决策模型 0.8B\u002F2B\u002F4B,基于阿里 Qwen3.5,七项评测 90.02% 平均准确率超过闭源 Jev 1.13.0 的 88.74%。4B 单次前向结构化决策,RTX 4090 上 289 token 三字段约 44 ms,权重开源。","## 决策模型赛道再添开源猛将\n\n决策模型(decision model)作为 LLM 的细分替代品类,正在被各家分头验证:Cloudflare 把 Clef 开源做评分、Perplexity 把 pplx-decider-v1-27b 放上 Hugging Face、TypeSafe 的闭源 Jev 在 JevBench 上长期作为基线。10 月 15 日,上海 AI Lab 的 Tong Zhu、Zhe Xu 等十位作者(联合上海交大、复旦、南洋理工)在 arXiv 挂出 Intern-Decision,以「多模态 + 单次前向」两个维度一起向前推。\n\n## 三档尺寸、Qwen3.5 基座、单一架构\n\nIntern-Decision 一次性放出 0.8B、2B、4B 三个档位,统一以 Qwen3.5(0.8B \u002F 2B \u002F 4B)为基座微调而来,视觉塔与投影层冻结。模型不再生成「自由文本 + JSON」,而是在 assistant 上下文里给每个字段预留一个占位符,直接读取原生 LM head 在该位置前的 logits,做合法候选上的 softmax,得到校准概率。所有字段一次前向搞定,没有迭代式 decode、没有 answer feedback。\n\n## 准确率单点超 Jev,校准数字更优\n\n- 七项评测套件平均 90.02%,Jev v1.13.0 同条件 88.74%(平均无加权)\n- 公开 Jevbench-Hard 111 题上,4B 直接答 82\u002F111(73.87%),温度缩放后 ECE 从 0.132613 降到 0.065267,Jev 同题 80\u002F111、ECE 0.094685\n- 96 题已知分布校准试点,4B 校准后 Brier 从 0.628 降到 0.550、ECE 从 0.213 降到 0.089,六个类别里五个优于 Jev\n\n## 推理速度、单卡可装\n\n固定 289 token、三字段请求(一个 choice、一个 boolean、一个 score)在单张 RTX 4090、BF16、SDPA 条件下:0.8B \u002F 2B 平均 33.28–33.98 ms、4B 44.16 ms,对应每秒 22.6–30.1 次串行请求。模型上下文窗口 32K,合法候选最多 62 个符号、单请求最多 16 个字段,这些是接口限制而不是泛化上限。\n\n## 一个值得抄的级联策略\n\n最值得工程团队抄的不是单点成绩,而是论文里做的置信度级联。把 111 道 Jevbench-Hard 喂给 4B,若置信度低于阈值就转给 Atria-Dawn-Preview 重新判一次,这套混合流程:\n\n- 4B 单跑 82\u002F111,73.87%\n- 置信度级联转 42 题到 Atria,合计 97\u002F111,87.39%(比 4B 多 15 题,比 Atria 单独少 2 题)\n- Atria 单跑 99\u002F111,89.19%\n- 混合估计耗时 140.16 秒,Atria 单跑 300.03 秒(论文报告值)\n\n也就是说:小模型负责明显对\u002F明显错的高置信请求,大模型只在不确定时救场,工程上最常见的小模型 + 大模型级联模式有了具体到数字的样板。\n\n## 所以呢\n\nIntern-Decision 把 Qwen3.5 重新做了一层专门面向结构化决策的后训练,4B 就已经在公开评测里压过闭源 Jev 1.13.0,这意味着:一,决策模型这个细分品类继续向开源 + 多模态演进,TypeSafe 的闭源定价策略将承受更大压力;二,Qwen 系开源生态在决策模型后训练底座上的吸引力再次被验证(Perplexity、Cloudflare、微软都选过 Qwen),下一波面向 agent 的二次微调很容易继续押在 Qwen3.5 上;三,这套架构给出的自回归掩码语言建模 + 单次前向多字段打分范式,可移植到任何带原生 LM head 的开源模型上,后训练数据与目标函数是真正的护城河。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.22131","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21,24,27],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":28,"name":29,"slug":29,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[31],{"id":32,"lang":33,"title":34,"summary":35,"content":36},"c7d1cfec-97a5-429e-8915-f974567d97d6","en","Intern-Decision: 4B tops Jev at 90.02%","Intern-Decision (Qwen3.5 0.8B\u002F2B\u002F4B): 90.02% across seven decision suites vs 88.74% for Jev; 4B serves 289-token requests in ~44 ms. Open weights.","## Decision Models Get Another Open-Source Contender\n\nThe decision-model category — a narrow LLM alternative for structured judgments — is getting crowded. Cloudflare open-sourced Clef, Perplexity put pplx-decider-v1-27b on Hugging Face, and TypeSafe's closed-source Jev has long served as the JevBench baseline. On October 15, ten authors led by Tong Zhu and Zhe Xu at Shanghai AI Lab, together with Shanghai Jiao Tong University, Fudan University, and Nanyang Technological University, posted Intern-Decision on arXiv, pushing the category forward on two axes: multimodality and single-pass scoring.\n\n## Three Sizes, Qwen3.5 Backbone, One Architecture\n\nIntern-Decision ships in three checkpoints — 0.8B, 2B, and 4B — all fine-tuned from matching Qwen3.5 backbones with the vision tower and projector frozen. Instead of emitting free-form text wrapped in JSON, the model reserves a placeholder token in the assistant skeleton for every decision field and reads logits off the native LM head at the position immediately before the placeholder, then runs a softmax over the legal candidate set to produce a calibrated probability. All requested fields are scored in a single causal forward pass; no iterative decoding, no answer feedback.\n\n## Single-Point Accuracy Beats Jev, Calibration Better Still\n\n- 90.02% unweighted average across seven evaluation suites, against 88.74% for Jev v1.13.0 on the same items.\n- On 111 public Jevbench-Hard items: Intern-Decision-4B answers 82\u002F111 (73.87%). After temperature scaling, ECE drops from 0.132613 to 0.065267; Jev hits 80\u002F111 with ECE 0.094685 on the same set.\n- On a 96-case known-distribution pilot, the calibrated 4B lowers Brier from 0.628 to 0.550 and ECE from 0.213 to 0.089; in five of six categories it beats Jev on both metrics.\n\n## Local Latency, Single-GPU Footprint\n\nFor a fixed 289-token request with three fields (one choice, one Boolean, one score), running BF16 with SDPA on a single RTX 4090: 0.8B and 2B average 33.28–33.98 ms; 4B averages 44.16 ms. The reciprocal rates correspond to roughly 22.6–30.1 serial requests per second. The context window tops out at 32K, the legal-candidate cap is 62 symbols, and the HTTP service permits up to 16 fields per request — those are interface limits rather than generalization ceilings.\n\n## A Confidence-Based Cascade Worth Copying\n\nThe most copyable engineering pattern isn't the headline number but the confidence-based cascade the paper studies end to end. Feeding the 111 Jevbench-Hard items to the 4B and routing low-confidence requests to Atria-Dawn-Preview:\n\n- 4B alone: 82\u002F111, 73.87%\n- Cascade (42 routed to Atria): 97\u002F111, 87.39% (+15 over 4B, –2 vs Atria-only)\n- Atria alone: 99\u002F111, 89.19%\n- Estimated total time: 140.16 s for the cascade, 300.03 s measured for Atria alone\n\nIn other words, a small model handles obvious cases, a larger one rescues the rest, and the engineering pattern is spelled out in concrete numbers rather than hand-waved.\n\n## So What\n\nIntern-Decision re-trains Qwen3.5 specifically for structured decisions, and the 4B already beats the closed-source Jev v1.13.0 on public benchmarks. Three things follow. One, the decision-model category is decisively moving toward open-source multimodal variants; TypeSafe's closed-source pricing will feel more pressure. Two, the Qwen family's attractiveness as a \"decision-model post-training base\" is reinforced again (Perplexity, Cloudflare, and Microsoft have all picked Qwen), so the next wave of agent-targeted fine-tuning will likely keep landing on Qwen3.5. Three, the autoregressive masked-language-modeling + single-forward multi-field scoring recipe is portable to any open-weight model with a native LM head — the training data and objective function are the real moat.","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06Z","2026-10-18T11:13:15.840739Z","2026-10-18T11:13:15.840749Z",true,"agent",243,[45,54],{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":41,"created_at":52,"modified_at":53},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":55,"tag_slug":55,"title_zh":56,"title_en":57,"intro_zh":58,"intro_en":59,"id":60,"is_active":41,"created_at":61,"modified_at":62},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":64},[65,70,75,80,85,90],{"id":66,"title":67,"news_slug":68,"published_at":69},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00",{"id":71,"title":72,"news_slug":73,"published_at":74},"4cc278cf-6b53-4722-81ac-d1909b50e6a9","Meta 开源 816B 模型，Mac 可本地跑","meta-muse-spark-open-816b-mac","2026-10-15T23:10:00+00:00",{"id":76,"title":77,"news_slug":78,"published_at":79},"9fcb5d74-315e-4d57-a7da-290fbfb9e91c","阿里 Qwen 发布弹性 RL 框架 QwenGyre,训练长时 Agent 提速 1.85x","alibaba-qwen-qwengyre-elastic-rl-xlong-agents","2026-10-14T00:00:00+00:00",{"id":81,"title":82,"news_slug":83,"published_at":84},"9c8731f9-e170-4855-93a2-b1705abb61c3","Muse Wonder发布:Meta把宝押在个人智能体上","meta-muse-wonder-agentic-model","2026-10-12T23:10:00+00:00",{"id":86,"title":87,"news_slug":88,"published_at":89},"5edbfca7-7c1a-4df3-8577-a0195408b827","RSI-Anything:世界模型陪跑,自我改进有了回滚键","rsi-anything-world-model-recursive-self-improvement","2026-10-12T21:20:00+00:00",{"id":91,"title":92,"news_slug":93,"published_at":94},"e40de2f9-9ee4-47f3-b1f1-7bf93a4870a9","一个动词翻转工具调用决策,LLM 内部向量现形","llm-tool-call-decision-vector","2026-10-09T13:10:00+00:00"]